agentscope.memory.temporary_memory module

Memory module for conversation

class agentscope.memory.temporary_memory.TemporaryMemory(embedding_model: str | Callable | None = None)

Bases: MemoryBase

In-memory memory module, not writing to hard disk

__init__(embedding_model: str | Callable | None = None) → None

Temporary memory module for conversation.

Parameters:

embedding_model (Union[str, Callable]) – if the temporary memory needs to be embedded, then either pass the name of embedding model or the embedding model itself.

add(memories: Sequence[Msg] | Msg | None, embed: bool = False) → None

Adding new memory fragment, depending on how the memory are stored :param memories: Memories to be added. :type memories: Union[Sequence[Msg], Msg, None] :param embed: Whether to generate embedding for the new added memories :type embed: bool

delete(index: Iterable | int) → None

Delete memory fragment, depending on how the memory are stored and matched :param index: indices of the memory fragments to delete :type index: Union[Iterable, int]

export(file_path: str | None = None, to_mem: bool = False) → list | None

Export memory, depending on how the memory are stored :param file_path: file path to save the memory to. The messages will

be serialized and written to the file.

Parameters:

to_mem (Optional[str]) – if True, just return the list of messages in memory

Notice: this method prevents file_path is None when to_mem is False.

load(memories: str | list[Msg] | Msg, overwrite: bool = False) → None

Load memory, depending on how the memory are passed, design to load from both file or dict :param memories: memories to be loaded.

If it is in str type, it will be first checked if it is a file; otherwise it will be deserialized as messages. Otherwise, memories must be either in message type or list

of messages.

Parameters:

overwrite (bool) – if True, clear the current memory before loading the new ones; if False, memories will be appended to the old one at the end.

clear() → None

Clean memory, depending on how the memory are stored

size() → int

Returns the number of memory segments in memory.

retrieve_by_embedding(query: str | list[Number], metric: Callable[[list[Number], list[Number]], float], top_k: int = 1, preserve_order: bool = True, embedding_model: Callable[[str | dict], list[Number]] | None = None) → list[dict]

Retrieve memory by their embeddings.

Parameters:

• query (Union[str, Embedding]) – Query string or embedding.

• metric (Callable[[Embedding, Embedding], float]) – A metric to compute the relevance between embeddings of query and memory. In default, higher relevance means better match, and you can set reverse to True to reverse the order.

• top_k (int, defaults to 1) – The number of memory units to retrieve.

• preserve_order (bool, defaults to True) – Whether to preserve the original order of the retrieved memory units.

• embedding_model (Callable[[Union[str, dict]], Embedding], defaults to None) – A callable object to embed the memory unit. If not provided, it will use the default embedding model.

Returns:

a list of retrieved memory units in specific order.

Return type:

list[dict]

get_embeddings(embedding_model: Callable[[str | dict], list[Number]] | None = None) → list

Get embeddings of all memory units. If embedding_model is provided, the memory units that doesn’t have embedding attribute will be embedded. Otherwise, its embedding will be None.

Parameters:

embedding_model – (Callable[[Union[str, dict]], Embedding], defaults to None): Embedding model or embedding vector.

Returns:

List of embeddings or None.

Return type:

list[Union[Embedding, None]]

get_memory(recent_n: int | None = None, filter_func: Callable[[int, dict], bool] | None = None) → list

Retrieve memory.

Parameters:

• recent_n (Optional[int], default None) – The last number of memories to return.

• filter_func – (Callable[[int, dict], bool], default to None): The function to filter memories, which take the index and memory unit as input, and return a boolean value.